Free guide Checked November 17, 2026. About 15 minutes to read. Parts change fast, so we list no prices; check today’s price at your store.
Why build your own AI computer?
Most people use AI by renting it: a monthly subscription for a chat app, another for pictures, another for writing help. Every month you pay, and your words travel to someone else’s computer. Owning is different. You buy a computer once, download free AI models, and they run in your own home. Your questions and files stay with you, nothing stops working when a company changes its prices, and it keeps working when the internet is down.
It is not magic, and it is not for everyone. The biggest cloud models are still stronger than anything you can run at home, and a good AI computer costs real money up front. But for everyday writing, summaries, homework help, private notes, pictures and learning how AI works, a home AI computer is very capable. This guide shows you how to build one without wasting money.
The one thing that matters most: video memory (VRAM)
An AI model is a very large file of numbers. To answer quickly, the whole model has to fit inside your graphics card’s own memory, called VRAM (video memory). If it does not fit, the computer falls back to the slower main processor and answers crawl. So the most important number on any AI computer is how much VRAM the graphics card has. If you remember one thing from this guide, remember that.
| Model size | Roughly how much VRAM | What it feels like |
|---|---|---|
| Small (3–8 billion parameters) | 6–8 GB | A good helper for writing, summaries and simple questions |
| Medium (12–14B) | 10–12 GB | Noticeably smarter, still quick |
| Large (27–32B) | 20–24 GB | Strong at reasoning and code |
| Very large (70B) | 40–48 GB | Needs two cards or a workstation card |
Rule of thumb for the common “4-bit” (Q4) versions of models: about 0.6 GB of VRAM per billion parameters, plus 1–2 GB for the conversation. These are approximations, not guarantees; longer conversations need more memory.
“Parameters” are the numbers inside a model; more parameters usually means a smarter but bigger model. “Quantized” or “4-bit” means the model was squeezed to use less memory with only a small loss in quality. Almost everyone running AI at home uses these squeezed versions.
Every part, in plain words
1. Graphics card (GPU): the engine
- NVIDIA is the easiest choice today. Almost every AI program supports it out of the box through NVIDIA’s CUDA software.
- AMD Radeon cards also work with popular apps like Ollama and LM Studio on supported models, and often give more VRAM for the money. Expect a little more setup.
- Apple Mac with lots of “unified memory” is another path: quiet and simple, very good for chatting with large models, slower for images and video, and not upgradeable.
2. Memory (RAM)
32 GB is the minimum we suggest. 64 GB or more if you want to run large models partly on the main processor or keep many apps open. RAM is cheap insurance compared with the graphics card.
3. Storage (SSD)
A fast NVMe SSD. Models are big files, from about 2 GB to over 40 GB each, so start with 1–2 TB. Choose 4 TB if you will collect many models or make video. A second cheap drive for backups is a good idea.
4. Processor (CPU)
It matters less than you think. Any modern 6–8 core processor from AMD or Intel is fine for a single-graphics-card build. Spend the difference on VRAM.
5. Power supply (PSU)
Do not save money here. Buy a quality unit (80 Plus Gold or better) that meets the card maker’s recommendation. Top cards draw a lot of power: NVIDIA lists the GeForce RTX 5090 at 575 W for the card alone and recommends a 1000 W system power supply.
6. Cooling and case
Big cards are long and hot. Pick a case with good airflow (a mesh front) and check that the card length fits. Quiet fans matter if it lives in your living room. AI work keeps the card busy for minutes at a time, much like a long game.
7. Motherboard
One full-size PCIe x16 slot is enough for one card. For two cards you need a board with two well-spaced slots and enough lanes; that is a “pro” build.
Three example builds
Pick by what you want to run, not by the biggest number. Model and card names are examples, not endorsements, and we earn nothing from them.
Try AI at home
- GPU: 12–16 GB VRAM (e.g. RTX 3060 12 GB, RTX 4060 Ti 16 GB, RTX 5060 Ti 16 GB, Radeon RX 7600 XT 16 GB)
- RAM: 32 GB
- SSD: 1–2 TB NVMe
- PSU: 650–750 W, Gold (check the card maker)
Runs small and medium chat models well, and image generation.
Daily AI workhorse
- GPU: 24–32 GB VRAM (e.g. used RTX 3090 24 GB, RTX 4090 24 GB, RTX 5090 32 GB, Radeon RX 7900 XTX 24 GB)
- RAM: 64 GB
- SSD: 2–4 TB NVMe
- PSU: 850–1000 W+, Gold (follow the card maker)
Runs large (about 30B) models comfortably, fast images, some video.
Home AI lab
- GPU: 48 GB+ total: two 24–32 GB cards, or one workstation card (e.g. RTX PRO 6000 Blackwell, 96 GB)
- RAM: 128 GB+
- SSD: 4 TB+ NVMe
- PSU: 1300–1600 W; check your home circuit
Runs very large (70B-class) models and heavy video work.
Step by step
- Write down what you want to do. Chatting and writing help? Making pictures? Running the biggest models? Your answer decides the graphics card, and the graphics card decides everything else.
- Pick the graphics card first. Use the VRAM table below. Buy the most video memory your budget allows, from a store with a good return policy.
- Pick a power supply the card maker recommends. Every card has a “recommended system power” on its spec page. Choose a quality unit (80 Plus Gold or better) that meets or beats it.
- Check the case fits. Look up the card’s length and thickness and the case’s “maximum GPU length”. Mesh fronts keep things cooler and quieter.
- Choose the rest: processor, motherboard, RAM, SSD. Any modern 6–8 core processor, a board that matches it, 32–64 GB RAM and a 1–2 TB NVMe SSD.
- Assemble it, or let a shop do it. Follow the motherboard manual page by page. Many local computer shops will assemble a PC for a small labour fee if you would rather not.
- Install the operating system and the graphics driver. Windows 11 or Ubuntu Linux. Then the newest driver from NVIDIA or AMD, straight from their website.
- Install LM Studio or Ollama and download a first model. Pick a model size from the table that fits your VRAM. After it downloads, it works with no internet.
- Test it, then keep it healthy. Check the model runs on the graphics card (not the processor), watch temperatures for the first week, and keep drivers updated.
The free software
Your first hour with the new PC
- Update everything. Run Windows Update (or
sudo apt update && sudo apt upgradeon Ubuntu), then install the newest graphics driver. - Check the card is seen. On NVIDIA, open a terminal and type
nvidia-smi. You should see your card’s name and its memory. - Install Ollama from ollama.com/download and run a small model:
ollama run llama3.2. Ask it something simple. - Check it is using the graphics card. In a second window type
ollama ps. Ollama’s own documentation says “100% GPU” means the model is fully on the graphics card. If you see CPU, pick a smaller model. - Try a bigger model that still fits your VRAM, and compare. This is how you learn what your machine can do.
Troubleshooting
| What you see | Most likely cause | What to do |
|---|---|---|
| The model answers very slowly, one word every few seconds | The model does not fit in video memory, so it spilled onto the processor | Choose a smaller model or a smaller “quantized” version (for example Q4). In Ollama, run ollama ps: it should say 100% GPU. |
| The PC turns off when the AI starts working hard | Power supply too small or poor quality | Check the card maker’s recommended system power. Use the separate power cables the card needs, not one split cable. |
| The app says it cannot find a GPU | Old or missing graphics driver | Install the newest driver from nvidia.com or amd.com, restart, and try again. |
| Fans are loud and the card is very hot | Poor airflow in the case | Make sure front fans pull air in and rear/top fans push it out. Leave space behind the PC. Clean dust every few months. |
| “Out of memory” error | Model plus conversation too big for VRAM | Use a smaller model, shorten very long conversations, or lower the context length setting. |
| The card does not fit | Card longer or thicker than the case allows | Check the “max GPU length” in the case manual before buying. Some cases let you remove a drive cage. |
Ahmad’s build Photos coming, in progress
Ahmad, who runs ABUZ8, is photographing his own AI workstation. The real photos and the full parts list will go here. Nothing below is a stock photo pretending to be his.
Questions people ask
Do I need to know how to code?
No. LM Studio is a normal app: pick a model, press download, chat. Code only matters if you later want to build your own agent.
Can I use the computer I already have?
Often, yes. Any PC with an NVIDIA or AMD graphics card that has 8 GB or more of video memory can run small models. Try it before you buy anything new.
Can I use a laptop?
Yes. A gaming laptop with an 8 GB+ NVIDIA card runs small models. It will be louder and slower than a desktop and you cannot upgrade the graphics card later.
What about a Mac?
Macs with Apple chips share their memory between the processor and graphics, so a Mac with a lot of memory can load surprisingly large models quietly. They are slower for images and video and cannot be upgraded after purchase.
Is one big card better than two smaller ones?
For a first build, yes. One card is simpler, quieter and needs a smaller power supply. Two cards make sense only when you need more total video memory than any single card you can afford.
Is a used graphics card safe to buy?
Many people buy used cards for AI. Buy from a seller with returns, test it under load the first day, and check the fans and temperatures.
How much does it cost?
Prices move every week, so we do not list them. Price the graphics card first; it is usually the biggest single cost. Then add the power supply, RAM and SSD.
Is my data private?
When the model runs on your own computer, your words and files stay there. Extra apps and plugins can still connect to the internet, so check their settings.
Sources
- NVIDIA GeForce RTX 5090 specifications (575 W card power, 1000 W recommended system power, 32 GB GDDR7)
- Ollama FAQ: ollama ps and the 100% GPU column
- Ollama download
- LM Studio
- llama.cpp (quantized model formats)
Next: turn this PC into an always-on AI server for the whole house.
Build my AI brain →Get started (optional sign-up)